Make Something AI Wants

We've spent thirty years learning how to build products for humans. Not just in the obvious ways — the button colors, the onboarding flows, the reassuring copy — but in subtler ones. We built trust through brand. We acquired customers through ads. We retained them through habit and switching costs. The whole edifice of modern consumer software rests on one assumption: somewhere in the loop, there's a person who can be charmed, persuaded, or confused.

That assumption is starting to break.

Gartner thinks AI agents will intermediate $15 trillion in B2B spending by 2028. Maybe they're wrong — analysts are often wrong about timing. But the direction seems hard to argue with. Already, companies are raising money explicitly to build banks for AI agents. Sean Neville, who co-created USDC, left that to start Catena Labs on the thesis that "there won't be any humans or businesses executing financial transactions directly. There will only be agents." Whether that happens in five years or twenty, it's pointing at something real.

The interesting question isn't whether agents will become major customers. It's what changes when your customer can't be charmed.

A human customer evaluates a product through a cloud of signals: the design, the brand, whether their colleagues use it, whether the website feels premium or cheap. Most of those signals have nothing to do with whether the product actually works. But they're surprisingly useful proxies. Over millions of years of social evolution, humans got pretty good at reading trustworthiness from surface cues.

Agents don't have any of that. They evaluate based on documentation quality, pricing clarity, API reliability, and whether similar agents have chosen the same product. They can't be distracted by an ad. They're not impressed by a landing page. They don't respond to a clever sales pitch.

This means agents are, in a weird way, better customers than humans. They actually measure whether your product does what you say. You can fool a human long enough to take their money. You probably can't fool an agent.

Y Combinator's partners noticed this when looking at which tools their AI-native startups were gravitating toward. Agents were overwhelmingly picking Supabase for databases — not because of brand or marketing, but because its documentation was the most machine-readable. The phrase one of them used was: "Documentation is the new front-end." It sounds almost too strange to be true. I think it is true.

In Patrick O'Brian's novels, captains always try to get upwind of their opponents. If you're upwind, you decide when and whether to fight. You control the engagement.

Some companies are getting upwind of the agent economy right now, without fully realizing it. Stripe is probably the best example. They co-developed an Agentic Commerce Protocol with OpenAI that lets agents check out directly inside ChatGPT. That's not charity — they understand that whoever builds the rails for agent payments ends up in an extraordinary position. Being the financial layer for agent commerce might be more valuable than being the financial layer for human commerce, because agent commerce won't have the friction that slows down human spending.

The interesting question is whether you have to be Stripe to play this game. I don't think you do. In fact, I suspect the advantage goes to small, focused companies right now, because large companies are too distracted defending their existing human-facing products to optimize hard for agents.

A small startup can make a decision that sounds almost too obvious: make something AI wants. Design every surface of your product for machine consumption. Write documentation that reads well to a language model. Build pricing that an agent can evaluate without talking to a sales rep. Expose an API that doesn't require navigating a UI built for humans.

There will be objections. Andrej Karpathy said recently that current agents are "slop" and won't really work for a decade. He might be right. But the cost of building agent-legible products is low, and the upside if he's wrong is large. This is the startup equivalent of building near a railroad before the railroad exists. It looks premature. That's usually when you want to do it.

Here's the flip that makes this interesting.

Building for agents isn't the opposite of building for humans. It's what building for humans was always supposed to be. Clear documentation, reliable APIs, honest pricing — these are things your human customers also want. You've just been getting away without them because humans are easy to distract.

When I think about the startups we've seen succeed long-term, they almost all share one trait: they built products that worked so clearly you didn't need to sell them. Stripe's documentation is famous. Twilio's API is famous. These companies made something that developers would recommend unprompted, partly because explaining it was easy. That's not a coincidence.

Agents don't remove the challenge of building a great product. They just strip away the ability to hide behind marketing, and force you to compete on the thing that mattered all along.

Here's a useful test. Imagine your best customer is an AI agent reading your documentation at 2 AM, trying to decide whether to use your API or a competitor's. There's no sales rep to answer questions. There's no logo to reassure it. There's no clever onboarding to hook it. All it has is: does this do what I need, can I trust it to keep working, and how much does it cost?

If the honest answer to those questions makes you nervous, you have work to do.

Y Combinator's motto is "make something people want." It's a good compass because it's stateless. You don't have to remember a strategy. You just keep asking: do people actually want this?

Make something AI wants works the same way. It cuts through a lot of noise. Should you add more features, or make the existing ones more reliable? What does an agent care about — features or reliability? Should you invest in marketing or documentation? What does an agent read? Should you optimize for signups or API uptime? What does an agent measure?

The answer is usually the same: strip away the surface and serve the actual function. Be the thing that works, not the thing that looks like it works.

I'm not saying human users are going away — they're not, at least not soon. But I think companies that treat "agent-legible" as an afterthought are making the same mistake companies made in 2005 when they treated mobile as an afterthought. The timing looked uncertain then too. The direction wasn't.

The best time to make something AI wants is probably right now, while most of your competitors are still optimizing for humans who can be distracted.